Sampling-attention deep learning network with transfer learning for large-scale urban point cloud semantic segmentation

Abstract

Targeting the development of smart cities to facilitate the significant analysis of large-scale urban for construction and update. This research develops a new method named transfer learning-based sampling-attention network (TSANet) to act on 3D urban point clouds for semantic segmentation. The main contributions of this research are a segmentation model and a transfer learning protocol, where the segmentation model adopts the point downsampling–upsampling structure for time efficiency, the embedding method and an attention mechanism for point cloud feature processing, and the transfer learning protocol is employed to reduce the data requirements and labeling efforts by using prior knowledge for practical application. In addition, a focal loss is designed to assist the model for feature extraction and learning with handling data imbalance. To demonstrate the efficiency and effectiveness of the developed method, a realistic point cloud dataset of Cambridge and Birmingham cities is utilized as a case study. The results demonstrate that (1) the developed method has promising performance with overall accuracy (OA) of 0.9133 and Mean Intersection over Union (MIoU) of 0.5588; (2) the proposed transfer learning protocol contributes to the core model performance by fully combining accuracy and time efficiency, offering a 74.91% improvement in time efficiency; (3) the developed TSANet outperforms other state-of-the-art models, such as PointNet++ and DGCNN, by comparing the accuracy and time efficiency. Overall, the method developed in this research is capable of great potential as a practical application tool by presenting accurate, effective, and efficient results for semantic segmentation of large-scale 3D urban point clouds.